Yan Junjie started as an intern at Baidu, building MiniMax with $500 million, covering 200 countries and 236 million users worldwide. Its stock price surged nearly eightfold in two months, with a valuation surpassing HKD 300 billion; meanwhile, under chip restrictions, MoE architecture and low-compute, high-efficiency routes are becoming a microcosm of China’s AI breakthrough. This article is adapted, compiled, and written by ForesightNews from Lin Wanwan.
(Background: Burning $705 monthly with zero revenue—this is a normal outcome for AI agent startups.)
(Additional context: Why did Trump call on multiple countries to defend the Hormuz Strait, yet have no allies backing him?)
Table of Contents
Toggle
- A bet made over a year earlier
- A peculiar shareholder chart
- 385 people and 1% of the money
- Great mountains can be crossed
In 2014, Baidu Research brought in an intern, a PhD from the Institute of Automation, Chinese Academy of Sciences, from a small town in Henan. He calculated his ideal post-graduation destination: IBM, coding in Java, with a yearly salary of 280,000 RMB.
Spring 2026, a tool called OpenClaw exploded globally, powered by large foundational models supporting developers building lobster-like agents. One model, fast and cheap, processed 1.44 trillion tokens in a week on OpenRouter, topping all platforms.
This model is called M2.5, and the company is MiniMax.
Two months after going public, its stock soared from HKD 165 to HKD 1,300, with a market cap breaking HKD 300 billion, despite annual revenue under $80 million.
The creator of MiniMax is the same intern from twelve years ago—Yan Junjie.
A bet made over a year earlier
Spring 2021, Yan returned to his hometown in Henan for the New Year and visited his grandfather.
His grandfather said he wanted to write a memoir, to record his 80 years of life. But he couldn’t type, nor organize his stories well. After a few attempts, he gave up.
Yan had been in AI for over a decade. At that moment, he suddenly realized that what he had built—though it had helped many companies—was of no use to an elderly person wanting to write a memoir.
This detail was later repeatedly cited, carrying a bit of an inspirational story flavor. But it truly explains his motivation: to make AI accessible for ordinary people. This obsession later drove a series of counterintuitive decisions.
At the end of 2021, he left SenseTime.
Timing was critical. SenseTime was preparing for a Hong Kong IPO; he was VP, Deputy Director of the Research Institute, CTO of the Smart City Business Group. He left at one of the company’s most valuable moments—before the IPO and wealth realization.
ChatGPT was released only in November 2022.
MiniMax was founded in December 2021.
This timing gap laid the foundation for everything later. Yan later said that if he hadn’t started early, in an environment where “star researchers and big tech AI backgrounds are more favored,” MiniMax wouldn’t have been able to compete.
His parents are ordinary folks. He studied high school in a county town, then math at Southeast University, later earned a PhD at the Institute of Automation, CAS, did a postdoc at Tsinghua, and joined SenseTime—rising step by step, with no overseas background or prominent connections.
During his internship at Baidu, he interacted with Yu Kai from Horizon Robotics. Yu Kai later said that academic ability can be trained, but those who can engineer AI into real products are rare. Yan Junjie is one of them.
After joining SenseTime, he spent seven years rising from intern to VP. In 2018, with limited staff, he led the development of an “All for One” model algorithm, surpassing CloudWalk and Yitu in bidding competitions, winning industry first. He was praised for “rapidly reading papers, focusing only on essentials, ignoring clichés.” This efficiency later became the company culture at MiniMax.
He named the company MiniMax, inspired by the minimax algorithm from game theory, developed by von Neumann.
He explained that decision-making should first guard against worst risks, then choose the relatively optimal solution.
A peculiar shareholder chart
In December 2021, MiniMax completed an angel round, raising $31 million, with a pre-money valuation of $170 million. Investors included MiHoYo, IDG, Hillhouse, and Yunqi.
MiHoYo’s investment was special. Yan and MiHoYo Chairman Liu Wei are close; Liu Wei invested at the angel stage, and he remains a non-executive director on MiniMax’s board.
MiHoYo is also a client—its NPC dialogue and story generation in games use MiniMax’s models.
After the angel round, a small incident occurred.
In March 2023, Silicon Valley Bank declared bankruptcy. At that time, all of MiniMax’s funds were in that bank. It was the riskiest moment early on—money lost, and the investment environment was chaotic. But they survived, and two months later secured Series A funding of $257 million, with a valuation of $1.157 billion.
Subsequently, the list of investors grew more impressive. Alibaba entered, Tencent joined, Sequoia Capital followed. Before going public, they completed seven rounds of financing, totaling nearly $1.5 billion, with a valuation of $4.2 billion. After IPO, Alibaba held 12.52%, making it the largest external shareholder.
Yan Junjie has a habit in early financing: only meet with top-tier investment firms. He met Sequoia’s Shen Nanpeng, and Hillhouse’s Zhang Lei.
But there’s one person on the shareholder chart worth highlighting: Yuan Yeyi.
Born in 1994, she studied electrical engineering at Johns Hopkins University, minored in economics and mathematics. Graduated in 2017, she joined SenseTime, focusing on financing and strategic investments. A year later, she became CEO Xu Li’s executive assistant and director of strategy. She deeply participated in SenseTime’s journey from early days to Hong Kong IPO.
In 2021, she and Yan Junjie started a company together.
An investor commented that she is “sharp, commanding, highly capable, with a maturity beyond her years.” Her role with Yan is clear: one defines the technical vision, the other turns that vision into money and resources. Yan can dive deep into technology, even with shaved hair, but the market, capital, and globalization are Yuan’s battleground.
On the day of the IPO, they stood together on the same stage. Yuan Yeyi, 31, valued at over HKD 40 billion.
385 people and 1% of the money
At IPO, MiniMax had 385 employees, average age 29.
From founding to September 2025, the company spent about $500 million. OpenAI spent between $40 billion and $55 billion in the same period.
This contrast is absurd. Using less than 1% of the opponent’s money, they built a globally leading multimodal company. Cost-saving was a result; the real reason was their mastery of AI. 80% of their code was generated by AI, internally called “interns,” with permissions high enough to access code repositories and modify online environments. They chat with these “interns” in Feishu, review code, and deploy directly.
This efficiency made MiniMax’s per capita output abnormally high.
Product-wise, they adopted a full multimodal approach from the start: language, video, speech, music—pushing all four simultaneously. While others focused on ChatGPT-style dialogue, Yan emphasized multimodal fusion. He believes multimodality is the fundamental premise for continuous intelligence improvement; without it, the next-gen models won’t have a chance.
In summer 2023, he made a more radical decision.
Allocate 80% of compute and R&D resources to MoE (Mixture of Experts).
At that time, domestic mainstream was still iterating dense models; MoE was considered “cutting-edge but immature.” Yan’s logic was simple: to serve tens of millions or hundreds of millions of users, the cost and latency of token generation with dense models are unsustainable. Without MoE, scale is impossible—everything else is pointless.
Early 2024, MiniMax released China’s first MoE large model.
In products, they didn’t compete in the domestic market. They launched Xingye and Talkie for C-end, one domestically, one overseas, focusing on AI companionship; Hailuo AI for video generation, which became the global leader in video generation app monthly active users for six consecutive months in late 2024.
Current figures: 236 million users, covering 200 countries and regions, with 73% overseas revenue. 214,000 enterprise clients and developers; models deployed in Google Vertex AI, Microsoft Azure, AWS. Notion’s first open-source model choice is also MiniMax.
In February, ARR surpassed $150 million. The M2 series’ daily token consumption is six times that of December last year, with a growth of over ten times in programming-related directions.
This explains why the market is willing to give a 200x sales multiple.
But some numbers need to be viewed separately.
In the annual report, C-end gross margin is 4.7%, B-end gross margin is 69.4%. 67% of revenue comes from C-end, but C-end barely contributes to gross profit. In Q4, C-end gross margin dropped to about 2.1%. Overall gross margin increased from 12.2% to 25.4%, mainly because B-end revenue rapidly increased in Q4, pulling up the overall figure.
This is an unresolved question.
Great mountains can be crossed
June 2025, MiniMax released the M1 model.
Yan Junjie posted on WeChat Moments:
“For the first time, I feel that great mountains are not insurmountable.”
Behind this statement lies a reality: the top AI models in China and the US may only differ by about 5% in capability, but that 5% allows overseas companies to dominate scenarios worth ten times more, charge ten times higher prices, and ultimately create nearly a hundredfold business gap. OpenAI’s latest valuation exceeds $700 billion. MiniMax’s market cap at listing is HKD 800 billion, less than $100 billion USD.
He made a judgment: in the future, there will be five top AGI companies globally, at least two from China, possibly one capable of leading.
After going public on January 9, he appeared on January 19 at a symposium hosted by the Premier, alongside other AI giants, becoming the second AI large model founder after Liang Wenfeng of DeepSeek to attend.
Then, on March 2, the first annual report was released, causing a surge in Hong Kong stocks.
In the earnings presentation, Yan spent a long time discussing: MiniMax must transform from a “large model company” into an “AI era platform company.”
He proposed a platform value formula: intelligence density × token throughput. In the internet era, platforms are traffic gateways; in the AI era, platforms are companies that define intelligence boundaries and share in commercial benefits. Google is doing it, OpenAI is doing it, and they aim to do it too.
His competitors are dozens of times larger.
Listing in Hong Kong only pushed him to another battlefield. Quarterly reports, analysts, market cap pressures—these are worlds apart from coding. The secondary market doesn’t believe in sentiment; it only looks at numbers. Can the C-end story translate into gross profit? Can B-end growth be sustained? When will M3 arrive? These questions must be answered every quarter.
But looking at the bigger picture, MiniMax’s story is more than just a company.
In recent years, the US has tightened chip controls. A100, H100, H800—restricted. The logic is straightforward: cut off compute power, choke AI development.
China has taken a completely different route.
DeepSeek achieved near-H100 performance using H800. MiniMax spent $500 million to do what OpenAI spent hundreds of billions on. Yan’s 2023 gamble on MoE was because their limited hardware couldn’t support inference for hundreds of millions of users. M2.5 running continuously for an hour costs $1, which is one-twentieth of GPT-5’s cost. Innovations like hybrid attention, linear attention, CISPO algorithms are all forced out by necessity.
The chip restrictions aimed to widen the gap, but the actual effect is pushing Chinese AI companies onto a low-compute, high-efficiency evolution path.
Less money, fewer restrictions, fewer people—yet this has driven extreme engineering capabilities and architectural innovation.
It’s similar to Huawei’s chip strategy: restrict one capability, and I’ll compensate in other dimensions. During this process, new things may emerge that you don’t have.
OpenAI now has over 4,000 people, spent $8 billion in 2025, and plans to invest $600 billion in compute by 2030. MiniMax has 385 people, with a total expenditure of $500 million.
Who will win remains unknown. But at least now, fewer people believe MiniMax will die.
That PhD from Henan who interned at Baidu in 2014 probably never imagined that twelve years later, he would stand at this position, backed by a whole national-level technological competition.
He chose to keep running.
